### CLOP for Noisy Black-Box Parameter Optimization

Posted:

**Thu Sep 01, 2011 10:32 am**Hi,

This is my paper for the Tilburg conference:

Title: CLOP: Confident Local Optimization for Noisy Black-Box Parameter Tuning

Abstract: Artificial intelligence in games often leads to the problem of parameter tuning. Some heuristics may have coefficients, and they should be tuned to maximize the win rate of the program. A possible approach consists in building local quadratic models of the win rate as a function of program parameters. Many local regression algorithms have already been proposed for this task, but they are usually not robust enough to deal automatically and efficiently with very noisy outputs and non-negative Hessians. The CLOP principle, which stands for Confident Local OPtimization, is a new approach to local regression that overcomes all these problems in a simple and efficient way. It consists in discarding samples whose estimated value is confidently inferior to the mean of all samples. Experiments demonstrate that, when the function to be optimized is smooth, this method outperforms all other tested algorithms.

pdf and source code:

http://remi.coulom.free.fr/CLOP/

It makes no miracle: you'll have to play a lot of games to get really good parameters. But it is certainly much more efficient than any manual method you could use with bayeselo. It is also more efficient than any other algorithm I am aware of.

Compared to the old version of QLR, I solved all the unstability problems. I do not have a mathematical proof of convergence, but I am convinced it always work well, unless the maximum is at a discontinuity, which never happens in practice.

Comments and questions are welcome.

Rémi

This is my paper for the Tilburg conference:

Title: CLOP: Confident Local Optimization for Noisy Black-Box Parameter Tuning

Abstract: Artificial intelligence in games often leads to the problem of parameter tuning. Some heuristics may have coefficients, and they should be tuned to maximize the win rate of the program. A possible approach consists in building local quadratic models of the win rate as a function of program parameters. Many local regression algorithms have already been proposed for this task, but they are usually not robust enough to deal automatically and efficiently with very noisy outputs and non-negative Hessians. The CLOP principle, which stands for Confident Local OPtimization, is a new approach to local regression that overcomes all these problems in a simple and efficient way. It consists in discarding samples whose estimated value is confidently inferior to the mean of all samples. Experiments demonstrate that, when the function to be optimized is smooth, this method outperforms all other tested algorithms.

pdf and source code:

http://remi.coulom.free.fr/CLOP/

It makes no miracle: you'll have to play a lot of games to get really good parameters. But it is certainly much more efficient than any manual method you could use with bayeselo. It is also more efficient than any other algorithm I am aware of.

Compared to the old version of QLR, I solved all the unstability problems. I do not have a mathematical proof of convergence, but I am convinced it always work well, unless the maximum is at a discontinuity, which never happens in practice.

Comments and questions are welcome.

Rémi